Overview
- OpenAI is offering the cyber-focused variant only through its Trusted Access for Cyber program, with a model ID that resolves for allowlisted API keys and returns a not-found error for standard keys.
- Approved users get fewer blocks for defensive jobs such as writing YARA detection rules, re-scoring known software flaws for their own stacks, studying malware behavior, and drafting incident reports from large log exports, and OpenAI says this preview eases those tasks rather than adding raw hacking power.
- The model refuses offensive outputs like shellcode, weaponized exploit proof-of-concept code, and command-and-control setup, and testers report a higher refusal rate even on some legitimate penetration-testing prompts.
- Independent checks show gains on narrow tasks compared with the standard GPT-5.5, with one public run solving 31 of 47 capture-the-flag challenges, though evaluators note such tests do not reflect performance against live, well-defended systems.
- Access is restricted to roughly a small allowlist through TAC, contrasting with Anthropic’s invite-only Mythos approach, and reported briefings to US and Five Eyes officials highlight rising oversight of who can use these tools and for what.